“Intelligent” controllers
on cheap and small
programmable devices
Cédric Joina,b,d, Frédéric Chaxelb and Michel Fliessa,c
Abstract— It is shown that the “intelligent” controllers which
are associated to the recently introduced model-free control
synthesis may be easily implemented on cheap and small pro-
grammable devices. Several successful numerical experiments
are presented with a special emphasis on fault tolerant control.
arXiv:1307.4841v1 [math.OC] 18 Jul 2013
Keywords— Model-free control, intelligent PID controllers,
small programmable devices, estimation, identification, noise
attenuation, fault tolerant control.
I. I NTRODUCTION
It is well know that the overwhelming majority of in-
dustrial control applications is based on PID controllers
(see, e.g., [1], [17], and the references therein). Those con-
trollers are often manufactured by numerous companies as
microcontrollers on cheap and small programmable devices,
like a Microchip1 PIC or a Freescale2 DSC for instance
This communication demonstrates that the recent model-free
control [6] and the corresponding intelligent PID controllers
[6] may also be implemented on such devices,
• thanks to the low power computation cost which they
require,
• although they obey to mathematical principles, which
are quite different from those of “classic” PIDs.
When compared to a PID regulator, its intelligent coun-
terpart contains one term more, which subsumes not only
the unknown structure of the plant but also the unknown
disturbances. This term, which is estimated online thanks to
recent parameter identification techniques ([8], [9]), is also
found in the linear ultra-local model which
• replaces the unknown global description of the plant,
• is continuously updated online via the same calcula-
tions,
• is of low order, which is most of the time equal to 1.
Remark 1.1: Let us emphasize moreover that the strange
ubiquity of classic PIDs was mathematically explained for
a AL.I.E.N. (ALgèbre pour Identification et Estimation Numériques), 24-
30 rue Lionnois, BP 60120, 54003 Nancy, France.
{cedric.join, michel.fliess}@alien-sas.com
b CRAN (CNRS, UMR 7039), Université de Lorraine, BP 239, 54506
Vandœuvre-lès-Nancy, France.
{cedric.join, frederic.chaxel}@univ-lorraine.fr
c LIX (CNRS, UMR 7161), École polytechnique, 91128 Palaiseau,
France. Michel.Fliess@polytechnique.edu
d Projet Non-A, INRIA Lille – Nord-Europe, France.
1 Let us quote Wikipedia: Microchip Technology is an American manu-
facturer of microcontroller, memory and analog semiconductors.
2 Let us quote again Wikipedia: Freescale Semiconductor is an American
company that produces and designs embedded hardware.
the first time, to the best of our knowledge, thanks to
intelligent PIDs [6].
This aim is fully justified by the following facts:
• Intelligent PIDs are much easier to tune than the classic
ones.
• They are robust with respect to most disturbances,
including quite strong ones.
• They permit straightforward fault accommodations.
• Many successful concrete applications were already
achieved in most various domains within a few years
(see the numerous references in [6]).
Our paper is organized as follows. Section II summarizes
some of the most important facts about model-free control
and its corresponding intelligent controllers. The implemen-
tation on small programmable devices is detailed in Section
III. Section IV describes some numerical experiments, with
a peculiar emphasis on fault tolerant control, i.e., on an
important topic in control engineering (see, e.g., [3], [16],
and the references therein). Several excellent simulations are
provided. Some concluding remarks may be found in Section
V.
II. M ODEL - FREE CONTROL : BASICS3
A. The ultra-local model
The unknown global description of the plant is replaced
by the ultra-local model
y (ν) = F + αu (1)
where
• the derivation order ν ≥ 1 is selected by the practi-
tioner;
• α ∈ R is chosen by the practitioner such that αu and
y (ν) are of the same magnitude.
Remark 2.1: Note that ν has no connection with the order
of the unknown system, which may be with distributed
parameters, i.e., which might be best described by partial
differential equations (see, e.g., [13] for hydroelectric power
plants).
Remark 2.2: The existing examples show that ν may
always be chosen quite low, i.e., 1 or 2. In almost all existing
concrete case-studies ν = 1. The only counterexample until
now where ν = 2 is provided by magnetic bearings [4] where
frictions are almost negligible.4
3 See [6] for more details.
4 See the explanation in [6].
Some comments on F are in order: Noise attenuation is achieved by multiplying both sides on
• F is estimated via the measure of u and y; the left by s−2 .5 It yields in the time domain the realtime
• F subsumes not only the unknown structure of the estimate
system but also any perturbation. Z t
6
Fest (t) = − 3 [(τ − 2σ)y(σ) + ασ(τ − σ)u(σ)] dσ
B. Intelligent PIDs τ t−τ
Set ν = 2 in Equation (1): where τ > 0 might be quite small. This integral may of
ÿ = F + αu (2) course be replaced in practice by a classic digital filter.

Close the loop via the intelligent proportional-integral- III. I MPLEMENTATION

derivative controller, or iPID,
R Let us remind that implementing controllers on small pro-
F − ÿ ∗ + KP e + KI e + KD ė
u=− (3) grammable devices is a well established topic in engineering
α (see, e.g., [11], [12], [18], [20], [21], [22], and the references
where therein).
⋆
• e = y − y is the tracking error,
• KP , KI , KD are the usual tuning gains. A. The iP device
Combining Equations (2) and (3) yields We first detail the implementation of the iP device, which
Z
is most of the time sufficient in practice.
ë + KD ė + KP e + KI e = 0
1) Device: The device is a microchip
where F does not appear anymore. The tuning of KP , KI , dsPIC33FJ128GP204 characterized by:6
KD is therefore quite straightforward. This is a major benefit • Architecture: 16-bit.
when compared to the tuning of “classic” PIDs (see, e.g., [1], • CPU speed (MIPS): 40.
[17], and the references therein). • Memory type: Flash.
Remark 2.3: If KI = 0 we obtain the intelligent • Program memory (KB): 128.
proportional-derivative controller, or iPD, We also utilize
F − ÿ ∗ + KP e + KD ė • two inputs with a 12-bit, 500 KSPS analog-to-digital
u=−
α conversion,
Set now ν = 1 in Equation (1):
• one 12-bit digital output coupled with an external
ẏ = F + αu (4) digital-to-anolog converter.
The loop is closed by intelligent proportional-integral con- The values of the input and output variables, which are
troller, or iPI, expressed in volts, are in the range [0, 3.3]. They can be
R connected however to an electronic amplifier stage. Let us
F − ẏ ∗ + KP e + KI e
u=− (5) add that our material is cheap: it costs less than 5 euros.
α Figure 1 displays the corresponding architecture:
If KI = 0, it yields an intelligent proportional controller, or
iP,
F − ẏ ∗ + KP e
u=− (6)
α

Remark 2.4: Equation (6) and the corresponding iPs are

most common in practice. This is again a major simplifica-
tion with respect to “classic” PIDs and PIs.
C. Estimation of F Fig. 1: Control scheme
F in Equation (1) is assumed to be “well” approximated
by a piecewise constant function Fest . According to the alge- 2) Simple calculations: A control iteration needs:
braic parameter identification developed in [8], [9], rewrite, • affectations: 19,
if ν = 1, Equation (4) in the operational domain (see, e.g., • conditions (if...then...else): 5,
[23]) • additions/subtractions: 14,
Φ
sY = + αU + y(0) • multiplications/divisions: 16.
s
where Φ is a constant. We get rid of the initial condition The computational power which is used is indeed very low.
d
y(0) by multiplying both sides on the left by ds : 5 See [5] for a theoretical explanation.
dY Φ dU 6 See http://www.microchip.com for more technical details, and
Y +s =− 2 +α also the prices.
ds s ds
B. Some extensions
Some new calculations are of course needed for iPIs,
iPIDs, and iPDs.
Remark 3.1: An iPD was employed only once, for mag-
netic bearings [4]. Note moreover that it was not necessary
until now to useR iPIDs in practice!
The integrals e appearing in iPIs and iPIDs may be dealt
via the classic trapezoidal rules (see, e.g., [2]). Less than 10
basic operations are used.
The derivative D(t) = KD ė in iPIDs and iPDs may be
obtained via
• a backward Euler difference scheme (see, e.g., [2])
e(tk ) − e(tk−1 )
D(tk ) = KD
tk − tk−1
• a low-pass digital filter for reducing the noise (see, e.g.,
[10]).
As above for the integration, less than 10 basic operations
are used.
Remark 3.2: With very noisy signals, more advanced
tools might be necessary for the differentiation (see [7], [15],
and [14]).
IV. N UMERICAL EXPERIMENTS
A. LabVIEW
As well known, LabVIEW greatly facilitates numerical
simulations in engineering and in science.7 LabView can be
programmed in order to control a real system by the use
of an input-output device (analogic and logic inputs and
outputs). Simple as well as advanced regulation strategies
may therefore be assessed. It may also be used for emu-
lating/simulating the behavior of a real system (see, e.g.,
[19]). Emulation is achieved in our experimental platform
via Labview and an input-output card (see Figure 2).
Fig. 2: LabVIEW details
Note that v becomes the true input control variable, where
−1.65 ≤ v ≤ 1.65. The non-negativity condition on u (see
Section III-A.1) is therefore dropped. Negative values for the
control variables are now possible. Figure 3 displays a full
description of our devices where the dsPIC is connected to
an interface from National Instrument.8
B. Two experiments
In both experiments the same iP is implemented where
α = 1 and KP = 1. The control variable is obtained thanks
to the dsPIC with a sample time equal to 0.1ms.
7 See http://www.ni.com/labview/whatis for more details.
8 See http://www.ni.com/white-paper/2732/en for details.
1) System 1: Consider the nonlinear stable system:
2(3v̇v 2 + v 3 ) = 0.5ÿ + 0.5ẏ − y
where v is the saturated control after offset (see Figure
2 for an explanation). The tracking performances, which
are presented in Figure 4, are quite good. They however
deteriorate with important change points. This is due to the
saturation of the control variable which is bounded.
Set
v̄ = πv
where
• π = 1 corresponds to the fault-free case,
• Figure 5-(c) shows the case 0 ≤ π  1, which
corresponds to a power loss of the actuator.
Figures 5-(a) and 5-(b) display an excellent fault tolerant
control even with violent faults.
2) System 2: Consider the nonlinear unstable system:
9v̇v 2 + v 3 = ÿ + ẏ + y
Close the loop with the same iP as above. Figure 6 displays
excellent tracking and fault accommodation performances.
V. C ONCLUSION
This communication has demonstrated that the intelligent
controllers, which are associated to model-free control, may
easily be implemented on cheap and small programmable
devices. Future applications will show that our academic nu-
merical experiments may be easily extended to more realistic
case-studies. It should lead to great industrial opportunities
for this new setting.
Acknowledgement: The authors thank the department of Génie
Électrique et Informatique Industrielle, Institut Universitaire de
Technologie Nancy-Brabois, Université de Lorraine, for its most
friendly help and its loan of some essential devices.
R EFERENCES
[1] K.J. Åström, T. Hägglund, Advanced PID Control, Instrument Soc.
Amer., 2006.
[2] K.E. Atkinson, An Introduction to Numerical Analysis (2nd ed.),
Wiley, 1989.
[3] M. Blanke, M. Kinnaert, J. Lunze, M. Staroswiecki, Diagnosis and
Fault-Tolerant Control, Springer, 2003.
[4] J. De Miras, S. Riachy, M. Fliess, C. Join, S. Bonnet, Vers une
commande sans modèle d’un palier magnétique, 7e Conf. Internat.
Francoph. Automatique, Grenoble, 2012. Available at
http://hal.archives-ouvertes.fr/hal-00682762/en/
[5] M. Fliess, Analyse non standard du bruit, C.R. Acad. Sci. Paris Ser.
I, vol. 342, pp. 797–802, 2006.
[6] M. Fliess, C. Join, Model-free control, Int. J. Control, 2013, DOI:
10.1080/00207179.2013.810345. Preprint available at
http://hal.archives-ouvertes.fr/hal-00828135/en/
[7] M. Fliess, C. Join, H. Sira-Ramı́rez, Non-linear estimation is easy, Int.
J. Model. Identif. Control, vol. 4, pp. 12–27, 2008. Preprint available
at
http://hal.archives-ouvertes.fr/inria-00158855/en/
[8] M. Fliess, H Sira-Ramı́rez, An algebraic framework for linear
identification, ESAIM Control Optimiz. Calc. Variat., 9: 151–168,
2003.
[9] M. Fliess, H. Sira-Ramı́rez, Closed-loop parametric identification for
continuous-time linear systems via new algebraic techniques, in eds.
H. Garnier and L. Wang, Identification of Continuous-time Models
from Sampled Data, Springer, pp. 362–391, 2008.
 Fig. 3: Our experimental setup with a dsPIC connected to an input-output card driven by LabVIEW

3.5

1.5
3

1
2.5

0.5
2

0
1.5

−0.5
1

−1
0.5

−1.5
0
0 50 100 150 0 50 100 150
time in s time in s

(a) Control (– blue) and control limits (- - red) (b) Setpoint (- - black) and output (– blue)
Fig. 4: Experimental results: fault-free case

[10] G.F. Franklin, M.L. Workman, D. Powell, Digital Control of Dynamic
Systems (3rd ed.), Addison-Wesley, 1997
[11] A.P. Godse, D.A. Godse, Microprocessors and Microcontrollers (6th
revised ed.), Technical Editions Pune, 2009.
[12] D. Ibrahim, Microcontroller Based Applied Digital Control, Wiley,
2006.
[13] C. Join, G. Robert, M. Fliess, Vers une commande sans modèle
pour aménagements hydroélectriques en cascade, 6e Conf. Internat.
Francoph. Automat., Nancy, 2010. Available at
http://hal.archives-ouvertes.fr/inria-00460912/en/
[14] D.Y. Liu, O. Gibaru, W. Perruquetti, Error analysis of Jacobi derivative
estimators for noisy signals, Numer. Algor., 58, 53–83, 2011.
[15] M. Mboup, C. Join, M. Fliess, Numerical differentiation with annihi-
lators in noisy environment, Numer. Algor., 50, 439–467, 2009.
[16] H. Noura, D. Theilliol, J.-C. Ponsart, A. Chamseddine, Fault-tolerant
Control Systems - Design and Practical Applications, Springer, 2009.
[17] A. O’Dwyer, Handbook of PI and PID Controller Tuning Rules (3rd
ed.), Imperial College Press, 2009.
[18] J.S. Parab, V.G. Shelake, R.K. Kamat, G.M. Naik, Practical Aspects
of Embedded System Design Using Microcontrollers, Springer, 2008
[19] A. St. Leger, A. Deese, J. Yakaski, C. Nwankpa, Controllable analog
emulator for power system analysis, Int. J. Electrical Power Energy
 3.5

1.5
3

1
2.5

0.5
2

0
1.5

−0.5
1

−1
0.5

−1.5
0
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
time in s time in s

(a) Control (– blue) and control limits (- - red) (b) Setpoint (- - black) and output (– blue)

1.1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2
0 20 40 60 80 100 120 140
time in s

(c) Multiplicative power loss

Fig. 5: Experimental results: faulty case

3.5

1.5
3

1
2.5

0.5
2

0
1.5

−0.5
1

−1
0.5

−1.5
0
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
time in s time in s

(a) Control (–,blue) and control limits (- -,red) (b) Setpoint (- - black) and Output (– blue)
Fig. 6: Experimental results: fault free case

Syst., 33, 1675–1685, 2011. 2011.

[20] C. Tavernier, Applications des microcontrôleurs PIC (4e éd.), Dunod, [21] J.W. Valvano, Embedded Microcomputer Systems – Real Time Inter-
 2

1.5 1.8

1.6
1

1.4

0.5
1.2

1
0

0.8

−0.5
0.6

0.4
−1

0.2

−1.5
0

0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
time in s time in s

(a) Control (– blue) and control limits (- - red) (b) Setpoint (- - black) and output (– blue)

1.1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2
0 20 40 60 80 100 120 140 160 180
time in s

(c) Multiplicative power loss

Fig. 7: Experimental results: faulty case

facing (3rd ed.), Cengage Learning, 2011.

[22] T. Wilmshurst, Designing Embedded Systems with PIC Microcon-
trollers: Principles and Applications, Elsevier, 2007.
[23] K. Yosida, Operational Calculus (translated from the Japanese),
Springer, 1984.